AI 回答样本中的品牌解释文本，如何做抽取和入库？

BRAND_PATTERNS = {
    "brand_a": ["品牌A", "BrandA", "A牌", "A品牌"],
    "brand_b": ["品牌B", "BrandB", "B牌"],
}

def locate_mentions(text: str, patterns: dict) -> list[dict]:
    """返回每个品牌在文本中的提及位置列表"""
    mentions = []
    for brand_key, aliases in patterns.items():
        for alias in aliases:
            start = 0
            while True:
                pos = text.find(alias, start)
                if pos == -1:
                    break
                mentions.append({
                    "brand": brand_key,
                    "alias": alias,
                    "position": pos,
                })
                start = pos + len(alias)
    return sorted(mentions, key=lambda x: x["position"])

def extract_explanation_segment(text: str, position: int, window_size: int = 300) -> str:
    """从提及位置向上下扩展，截取解释段落"""
    # 向上寻找段落边界
    para_start = max(0, position - window_size)
    prev_break = text.rfind("\n\n", 0, position)
    if prev_break != -1 and prev_break > para_start:
        para_start = prev_break + 2
    
    # 向下寻找段落边界
    para_end = min(len(text), position + window_size)
    next_break = text.find("\n\n", position)
    if next_break != -1 and next_break < para_end:
        para_end = next_break
    
    return text[para_start:para_end].strip()

CREATE TABLE brand_explanation (
    id BIGINT AUTO_INCREMENT PRIMARY KEY,
    task_id VARCHAR(64) NOT NULL,
    platform VARCHAR(30) NOT NULL,
    scene_type VARCHAR(30) NOT NULL COMMENT '场景分类',
    brand_name VARCHAR(100) NOT NULL COMMENT '被解释品牌',
    
    explanation_text TEXT COMMENT '抽取的解释段落原文',
    char_count INT DEFAULT 0 COMMENT '解释文本字符数',
    has_substantive_desc TINYINT(1) DEFAULT 0 COMMENT '是否包含实质描述',
    
    is_accurate TINYINT(1) DEFAULT NULL COMMENT '人工复核: 描述是否准确',
    is_complete TINYINT(1) DEFAULT NULL COMMENT '人工复核: 是否遗漏关键信息',
    risk_label VARCHAR(50) COMMENT '异常标记: 混淆/错误/遗漏/正常',
    
    raw_data_url VARCHAR(512) COMMENT 'COS原始回答链接',
    created_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP,
    
    INDEX idx_brand_task (brand_name, task_id),
    INDEX idx_risk (risk_label)
) ENGINE=InnoDB DEFAULT CHARSET=utf8mb4；

SELECT id, explanation_text
FROM brand_explanation
WHERE brand_name = '品牌A'
  AND (explanation_text LIKE '%品牌B%' OR explanation_text LIKE '%BrandB%')
  AND risk_label IS NULL；

def classify_risk(explanation: dict) -> str:
    risk_score = 0
    risk_tags = []
    
    if explanation["has_competitor_mention"]:
        risk_score += 2
        risk_tags.append("混淆")
    
    if explanation["fact_mismatch_count"] > 0:
        risk_score += 3
        risk_tags.append("错误")
    
    if explanation["missing_key_info_count"] >= 2:
        risk_score += 1
        risk_tags.append("遗漏")
    
    if risk_score >= 3:
        return "high_risk:" + ",".join(risk_tags)
    elif risk_score >= 1:
        return "low_risk:" + ",".join(risk_tags)
    return "正常"

SELECT 
    brand_name,
    platform,
    COUNT(*) AS total_explanations,
    SUM(has_substantive_desc) AS substantive_count,
    SUM(CASE WHEN risk_label = '正常' THEN 1 ELSE 0 END) AS accurate_count,
    AVG(char_count) AS avg_explanation_length,
    ROUND(SUM(CASE WHEN risk_label = '错误' THEN 1 ELSE 0 END) * 100.0 / COUNT(*), 1) AS error_rate
FROM brand_explanation
WHERE created_at >= '2026-06-01'
GROUP BY brand_name, platform
ORDER BY error_rate ASC；